ImaGene

ImaGene applies convolutional neural networks (CNNs) to detect and quantify natural selection in population genomic datasets.

Key Features:

• Deep Learning Approach: Employs convolutional neural networks (CNNs) as a supervised machine learning model to recognize patterns in genomic data.
• Data Representation: Transforms aligned genomic data from multiple individuals into images by stacking rows and encoding distinct alleles using different colors.
• Training on Simulated Data: Trains models on simulated datasets to learn signatures of positive selection.
• Detection of Selection Signatures: Identifies signatures of positive selection in genomes.
• Quantification of Selection Strength: Estimates selection coefficients and produces posterior distributions with confidence metrics.
• Image Sorting for Accuracy: Utilizes sorting of genomic images by rows and columns to enhance prediction accuracy.
• Demographic Model Considerations: Accounts for biases introduced by misspecified demographic models during training.
• Multiclass Classification for Continuous Variables: Applies multiclass classification techniques to estimate continuous variables such as the selection coefficient.
• Joint Inference Potential: Facilitates joint inference of mutation history and functional impact for mapping studies.

Scientific Applications:

• Evolutionary Framework Analysis: Identifies genomic sites targeted by natural selection to provide an evolutionary perspective on genetic loci implicated in phenotypes.
• Quantification of Selection Strength: Provides estimates of selection coefficients and uncertainty measures to characterize evolutionary pressures.
• Mapping and Functional Interpretation: Supports mapping studies by enabling joint inference of mutation history and functional impact relevant to molecular mechanisms of human phenotypes.

Methodology:

Transforms aligned genotype matrices into colored stacked images, trains supervised convolutional neural networks on simulated datasets, sorts genomic images by rows and columns to improve accuracy, considers demographic model misspecification during training, employs multiclass classification to estimate continuous variables like selection coefficients, and outputs posterior distributions with confidence metrics.